Canonical Correlation Analysis of Transient EMG Data for Multi-User Motion Intent Detection

Surface electromyography (sEMG) plays a relevant role in pattern recognition problems mainly related to upper limb motion. Lately, particular attention is given to the motion intent detection (MID) which involves transient epoch of the sEMG signal to decode and predict the movement that is going to be performed, which is a key aspect in the development of modern assistive devicesand myoelectric prosthesis. However, results achieved up to now are only related to an intra-subject scenario, while the user-independent case remains less investigated. For this reason, the present study focuses in defining an approach to face the multi-user MID problem, taking advantage of the least square canonical correlation analysis (LS-CCA). Considered data for this study belong to a publicly available dataset and they contain four shoulder movements of eight subjects. In the defined framework, the LS-CCA is used to create a common unified-space where features related to different subjects are maximally correlated among them. Then, the classification is performed through a SVM model. Performance of the classifier was evaluated also increasing the number of calibration trials and training subjects. Moreover, two window lengths are considered for feature extraction, i.e. 150 ms and 50 ms. Obtained results showed an improvement brought by the LS-CCA with an increasing trend of the classification accuracy when more training subjects and more calibration trials are considered; their values passed from 25% without LS-CCA up to 74% with LS-CCA.